wpferrell/qwen3-8b-bigsmall

Qwen 3 8B — Lossless Compressed

15.26 GB → 10.08 GB (34% smaller). Bit-identical weights. Drop-in replacement.

Use it in 2 lines

pip install bigsmall

from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("wpferrell/qwen3-8b-bigsmall")

It works exactly like loading the original model. No code changes needed.

Size comparison

	Size
Original (Qwen/Qwen3-8B)	15.26 GB
This compressed version	10.08 GB
Saved	5.18 GB (34%)

What "lossless" means

Every weight is mathematically identical to the original model.

• Not quantized. Quantization rounds weights and changes model behaviour.
• Not pruned. Pruning removes parts of the model.
• Bit-for-bit identical. md5 is verified on every tensor at decompression.

Low-VRAM streaming

from bigsmall import BigSmallStreamingModel

model = BigSmallStreamingModel.from_pretrained(
    "wpferrell/qwen3-8b-bigsmall",
    device="cuda",
    lru_max_vram_gb=2.0,
)

Uses up to ~12× less VRAM than standard loading by streaming layers on demand.

Decompress to safetensors

pip install bigsmall
bigsmall decompress wpferrell/qwen3-8b-bigsmall -o qwen3-8b-bigsmall/

Original model

This is a lossless-compressed copy of Qwen/Qwen3-8B. All credit to the original authors. The weights are unchanged.

Want to compress your own model?

pip install bigsmall
bigsmall compress my-model/ -o my-model.bs

See github.com/wpferrell/Bigsmall for the full docs.

License

• Model weights: apache-2.0 — same as Qwen/Qwen3-8B.
• BigSmall format: Elastic License 2.0 — free for personal, research, and commercial use.
• Commercial SaaS licensing: wpferrell@gmail.com

Citation

@misc{bigsmall2026,
  title={BigSmall: Lossless Neural Network Weight Compression},
  author={Ferrell, Will},
  year={2026},
  doi={10.5281/zenodo.20279248},
  url={https://doi.org/10.5281/zenodo.20279248}
}

Requires

bigsmall >= 3.13.0 for the latest features. Earlier versions (>= 3.0.0) can still decode this model.